Since traffic on a physical loop flows through all nodes on the loop, including the connecting fiber or copper links, the two primary sources of latency on an arbitrated loop come from:
These latencies have been demonstrated to be the primary cause of degraded performance for applications such as implementing a higher layer SCSI protocol over a Fibre Channel Arbitrated Loop protocol.
FIG. 1 shows an FC-AL node with a 6-word FIFO designated by reference number 3. It has a TX port 1 and a RX port 2. Each FC unencoded 8-bit character is translated to 10 bits when encoded on the serial link via the 8-bit to 10-bit encoding as defined by the Fibre Channel Physical and Signalling Interface (FC-PH) standard, ANSI X3.230-199x. There are four characters per FC word. The 6 FC word FIFO is the cause of the 6 word delay between the time an FC word arrives on RX port 2 and is retransmitted on TX port 1.
FIG. 2 shows the relationship between data frames and primitives. A primitive (sometimes also referred to as a command herein) is a FC word which occupies the inter-frame spaces, and has special meaning for flow control and loop management.
The primitives relevant to the invention are shown in FIG. 3. The ARB and OPN primitives contain addressing information. The ARB contains the address of the arbitrating node designated in FIG. 3 as AL.sub.-- PA. The AL.sub.-- PA is duplicated in the fourth character. The OPN primitive contains the destination node address (AL.sub.-- PD) in the third character and the source node address (AL.sub.-- PS) in the fourth character. Fill words are ARB primitives or IDLE primitives which are used by FC-AL nodes to perform loop signalling. Therefore, all fill words are primitives, but not all primitives are fill words.
FIG. 4 shows a four-node Fibre Channel Arbitrated Loop having the prior art unaccelerated architecture. The output of one node is connected to the input of the subsequent node, and so on. The sum of the latencies of the nodes (each node's 6 word delay) plus the inter-node link propagation delays is referred to as the "system latency."
FIG. 5 shows a "loop tenancy," which is the hanshaking protocol (a protocol is a set of rules for communicating) which defines the traffic which occurs between nodes to acquire control of an FC-AL to allow two nodes to communicate and which occurs before the loop is relinquished by these two nodes so that other nodes can communicate. In other words, a loop tenancy protocol is carried out so that a source node and a destination node can acquire the loop for their exclusive use in a data transfer operation. Each node has a priority ranking which is used during a process called arbitration. Arbitration is a process to decide which of 2 or more nodes which are simultaneously requesting control of the loop will get control of the loop. In the loop tenancy protocol shown in FIG. 5, an arbitration occurs, followed by an open primitive (hereafter referred to as an OPN primitive or simply an OPN) which is transmitted by the winning node. An OPN primitive is addressed to a specific node (the destination node) which returns an RRDY primitive when it is ready to receive data. Receipt of the RRDY by the source node which initiated the OPN is followed by transmission of one or more data frames from the source node to the destination node. Transfer of data is followed by transmission of a close primitive (hereafter a CLS primitive or simply a CLS) which can be transmitted by either node. When the source node receives a CLS from the destination node (or transmits a CLS to the destination node), it relinquishes control of the loop.
There follows a more detailed discussion of each phase of the prior art loop tenancy protocol.
Arbitration or ARB Primitives:
A node knows when it has won arbitration when it sees an inbound ARB primitive containing it's own AL.sub.-- PA priority ranking. Algebraically small AL.sub.-- PAs are higher priority than algebraically large AL.sub.-- PAs. If a port wishes to arbitrate and it receives a lower priority arbitration, it subsitutes its own ARB, i.e., it transmits an ARB with its own AL.sub.-- PA. If it receives a higher priority ARB, it passes that higher priority ARB.
Open or OPN Primitives:
OPNs are passed by a receiving port if the destination address does not match the AL.sub.-- PA of the receiving port.
Permission to Send or RRDY Primitives:
RRDYs are returned by the OPN recipient, i.e., the destination node having the AL.sub.-- PA which matches the AL.sub.-- PD in an OPN primitive. Each RRDY received by the OPN initiator, i.e., the source node, gives the OPN initiator permission to transmit one data frame. In the zero BB.sub.-- Credit model, the OPN initiator may not transmit data until one or more RRDYs are received.
Close or CLS Primitive:
A CLS primitive may be initiated by either the source or destination node. If a node receives a CLS and did not originate a CLS, it must forward it. If both nodes originate a CLS simultaneously, since there is no addressing in a CLS, both nodes will believe the incoming CLS from the other node was their CLS, and both will close the loop concurrently.
Performance:
FIG. 6 shows a series of loop tenancies which make up the protocol of a higher layer SCSI Write operation superimposed upon a lower layer FC-AL protocol. There are four FC-AL loop tenancies in a SCSI write operation: one for sending the write command, one for acknowledging receipt of the write command by a disk drive, one for sending the actual write data, and one for acknowledging receipt of the write data by the disk drive. Each ARB and CLS must be passed through every node on the FC-AL in the prior art FC-AL protocol, while each data frame and OPN primitive must be passed through, on the average, half the nodes on the loop (since data frames and OPN primitives are not propagated by the destination node).
With 4 loop tenancies and an average of 3 round trips per tenancy, there are 12 round trips per SCSI write command. Multiply this by the number of nodes on the loop and again by 6 words delay per node to get 72.times.N word delays per SCSI write command where N is the number of nodes on the FC-AL.
For a fully configured, two-initiator loop (2 initiators or servers, and 62 disk drives per initiator, where each initiator and disk drive is a node), the above delay or loop latency calculation becomes 72.times.126=9072 word delays, or 9072.times.40=363 kbit delays per SCSI write command. At 1.062 Gbits/sec, this equals about 341 microseconds of link overhead per SCSI write command.
For disk drives which have around 500 microseconds of controller software and hardware overhead, having a fixed 341 microsecond delay due just to FC-AL overhead significantly reduces the number of SCSI operations per second per arbitrated loop.
Therefore, a need has arisen for a bridge product which can reduce loop latency as well as reduce contentions for the loop in an FC-AL protocol network.